Works withClaudeChatGPTClaude CodeCodexCursor
Your finance data, checked before your AI reads it.
Harnace checks every record in your ledger, your CRM and the workbook you trust, so the AI you already use counts the way your team does.
Without HarnaceIllustration. The runs are an illustration of what we expect, until a real run fills them.
3 answers in 10 runs
7 runs said $167,000, the ledger is missing the expansion. 2 runs said $24,000, an error in the CRM. 1 run said $80,000, the CRM’s deals against the ledger.
With HarnaceIllustration. The runs are an illustration of what we expect, until a real run fills them.
Not asked yet1 answer in 10 runs
10 runs said Timing, $24,000: the expansion’s service starts on 1 October.
HubSpot · CRM
The deal’s amount
$167,000
Counts from the day it closed
Rillet · Ledger
Lines in service
$143,000
Counts the expansion from 1 October
ARR workbook
Its line on the ARR tab
$143,000
Agrees with the ledger
$24,000 apart
ReadingRillet6 of 6
- Source596 contract lines
- Fieldmonthly fee, read as ARR
- DefinitionARR v1, from service start
- Calculationmonthly fee times twelve
- OwnerDana Whitfield, VP Finance
- Dateservice start, per line
ARR v1 · approved by Dana Whitfield, VP Finance, 11 Oct 2026
Claude’s answerHarnace connector · Read only
Timing, $24,000: the expansion’s service starts on 1 October
- Definition
- ARR v1, counted from service start
- Approved by
- Dana Whitfield, VP Finance, 11 Oct 2026
- From
- 6 contract lines in service on 30 September 2026
Without HarnaceAbout 179,000 tokens read, $0.42 an answer, about 45 seconds*
With HarnaceAbout 30,000 tokens read, $0.05 an answer, about 10 seconds*
Measured in our reader eval on Marlow with Claude Sonnet 5.5: the median read and time, and the average cost of an answer, across all twenty questions. Not this question’s own.
01 / 06
Your AI reads the raw exports. Ask it the same question ten times.
02 / 06
It gives three answers. Seven of ten say $167,000, and they’re wrong.
03 / 06
Each system is right by its own rule. Nothing in the files says which rule Marlow reports by.
04 / 06
Harnace reads every record first, against the definition your team approved.
05 / 06
Ask again. Ten runs, one answer.
06 / 06
The same answer every run, and you can see where it came from.
The figures on this page follow Marlow Systems, a company we invented to show how a finance team uses Harnace. Its people and numbers are made up.
01 / 06
Your AI reads the raw exports. Ask it the same question ten times.
02 / 06
It gives three answers. Seven of ten say $167,000, and they’re wrong.
Without HarnaceIllustration. The runs are an illustration of what we expect, until a real run fills them.
3 answers in 10 runs
7 runs said $167,000, the ledger is missing the expansion. 2 runs said $24,000, an error in the CRM. 1 run said $80,000, the CRM’s deals against the ledger.
03 / 06
Each system is right by its own rule. Nothing in the files says which rule Marlow reports by.
HubSpot · CRM
The deal’s amount
$167,000
Counts from the day it closed
Rillet · Ledger
Lines in service
$143,000
Counts the expansion from 1 October
ARR workbook
Its line on the ARR tab
$143,000
Agrees with the ledger
$24,000 apart
With Harnace
04 / 06
Harnace reads every record first, against the definition your team approved.
ReadingRillet6 of 6
- Source596 contract lines
- Fieldmonthly fee, read as ARR
- DefinitionARR v1, from service start
- Calculationmonthly fee times twelve
- OwnerDana Whitfield, VP Finance
- Dateservice start, per line
ARR v1 · approved by Dana Whitfield, VP Finance, 11 Oct 2026
05 / 06
Ask again. Ten runs, one answer.
Without HarnaceIllustration. The runs are an illustration of what we expect, until a real run fills them.
3 answers in 10 runs
7 runs said $167,000, the ledger is missing the expansion. 2 runs said $24,000, an error in the CRM. 1 run said $80,000, the CRM’s deals against the ledger.
With HarnaceIllustration. The runs are an illustration of what we expect, until a real run fills them.
Not asked yet1 answer in 10 runs
10 runs said Timing, $24,000: the expansion’s service starts on 1 October.
06 / 06
The same answer every run, and you can see where it came from.
Claude’s answerHarnace connector · Read only
Timing, $24,000: the expansion’s service starts on 1 October
- Definition
- ARR v1, counted from service start
- Approved by
- Dana Whitfield, VP Finance, 11 Oct 2026
- From
- 6 contract lines in service on 30 September 2026
Without HarnaceAbout 179,000 tokens read, $0.42 an answer, about 45 seconds*
With HarnaceAbout 30,000 tokens read, $0.05 an answer, about 10 seconds*
Measured in our reader eval on Marlow with Claude Sonnet 5.5: the median read and time, and the average cost of an answer, across all twenty questions. Not this question’s own.
Every record, from your systems to your AI.
Harnace reads your ledger, your CRM and the workbook you trust, checks every record and puts it in your company’s own terms, so the AI you already use counts the way your team does.
Records from your systems, Salesforce, Rillet, Paylocity for Finance, Navan, Rippling and your contracts and files, flow into Harnace. Harnace understands them, connects them and validates them, then hands trusted context to any AI tool or agent: Claude, ChatGPT, Codex, Claude Code, Cursor, your own agents, BI tools and the rest of your tools.
Less slop, more meaning.
The same twenty questions about Marlow, asked of the same AI two ways: once on its raw exports and once through Harnace. Each square is one answer. A filled square is right. An empty one is wrong.
Point at a column, or pick one, to see its question.
RightWrongA and B are two AI models, each asked five times.
Skills give you consistency, Harnace gives you auditability.
A skill teaches your AI a convention, and it applies it the same way every time. If the convention is wrong, the wrong answer arrives the same way every time too, and nobody can say which version produced last quarter’s number. Keep your skills. Harnace is what they read, so every answer can be checked.
Which definition was it?
Every definition is written down, numbered and dated. Once approved it’s frozen, and a change becomes the next version, so the one before stays on the record.
Who approved it?
Each definition names the person who owns it: at Marlow Systems, Dana Whitfield, VP Finance. A new version is in force only once a named person approves it, and before that Harnace checks it against the report the owner already trusts.
Is it one customer or two?
One customer turns up under different names, account numbers and ship-to addresses across your CRM and your ledger. Harnace matches them to one customer before anything is counted. At Marlow Systems, Ondine Studios is two accounts in HubSpot and one customer in Rillet, so its $188,000 of ARR is counted once.
What produced last quarter’s number?
Every figure your AI is given is logged with the definitions it was counted under and how each system’s fields were read, and it traces back to the records behind it.
Why don’t the systems agree?
Where your ledger, your CRM and the workbook disagree, Harnace shows the difference in dollars, with its cause.
One figure, as it is logged
- Figure
- ARR at 30 Sep 2026, $3,810,000
- Counted under
- ARR v1, today’s template, counted from service start
- Approved by
- Dana Whitfield, VP Finance, 11 Oct 2026
- Read from
- 234 contract lines in Rillet
- Checked against
- The ARR workbook Dana Whitfield trusts
Every figure, traced back to where it came from.
Your sources, the records they hold, the fields that matter, your definitions and what reads them, on one map. A difference shows where it starts, so a person can explain it or fix the record at its source.
Dana Whitfield, VP FinanceContext map
Today, 6 Oct- 23 current
- 5 waiting on you
The Context map in Harnace’s console, as Marlow’s finance team sees it.
Read only, and yours to take with you.
Harnace reads your systems with the narrowest access each one allows. What it holds for you stays yours. In the words of our data addendum: We do not sell it, and we do not use it, or let anyone use it, to train a model.
What stays yours, point by point
It reads. It doesn’t write.
Harnace changes nothing in your systems. Where a system offers read-only access, Harnace asks for that alone. Where it doesn’t, no path in our code writes to your ledger, your CRM or your payroll, and a test fails our build if one appears.
Each company, fenced
Every record carries its company, and the database itself keeps each company’s records apart, so one company can never read another’s. A proof checks it.
Keys stay sealed
The keys to your systems are encrypted where they’re kept and never shown to a model.
Documents are data
An invoice, an email or a file is read as data. Nothing written inside one is followed as an instruction.
Every read, on the record
Each read by a person or an app is logged under who made it. You can cut off any person or app in one click.
A ledger of every call
Every call Harnace makes is written to a ledger chained so that any edit to it shows.
Bring one report you trust. We'll show you where your systems disagree.
Three sources, one report.
Bring the report you trust and the three systems behind it, usually your ledger, your CRM and the workbook. We rebuild the report from those systems, explain every difference in dollars and connect your AI to the result. Sixty days.